AmpereOne A192-32X Benchmarks: 192 Core ARM Server Performance & Power Efficiency
In total 70 benchmarks were run for this initial performance testing of the AmpereOne A192-32X processor. When taking the geometric mean of all the processors that each successfully ran all of the benchmarks, above is the geometric mean of all those raw performance results. The AmpereOne A192-32X performed admirably in matching the performance of a dual Intel Xeon 6766E Sierra Forest configuration overall but behind the Xeon 6780E 2P configuration. The AmpereOne A192-32X in a single socket configuration was also behind the single EPYC 9654 Genoa / 9754 Bergamo / 9684X Genoa-X processors.
Going from the Ampere Altra Max M128-30 to AmpereOne A192-32X was 1.56x the performance overall. That's with 1.5x the core count in going from Ampere Altra Max 128 cores to 192 cores with the current top-end AmpereOne SKUs. As shown in the few single-threaded and lightly threaded workloads, there wasn't too much IPC uplift going to AmpereOne from Ampere Altra cores. With AmpereOne is also the move to DDR5 memory from DDR4 and other improvements.
As the Sierra Forest SKUs were choking on some of the heavy HPC benchmarks, above is also the geometric mean if omitting all of the Xeon 6766E / 6780E results to show the geometric mean of all the CPUs sans Sierra Forest that ran all of the benchmarks successfully. The AMD EPYC 9754 128-core Zen 4 processor was running at 1.11x the performance of the AmpereOne A192-32X overall even with the lower core count. With the HPC exposure, the generational uplift here going from the Ampere Altra Max to AmpereOne top-end parts was 1.72x the performance.
When looking at the CPU power consumption over the span of all the benchmarks carried out, the AmpereOne A192-32X on average was consuming 230 Watts with a peak of 401 Watts, up from a 166 Watt average for the Ampere Altra Max M128-30 with a 166 Watt average and 252 Watt peak. The most significant difference with AmpereOne was unfortunately the much higher CPU power consumption at idle and lightly threaded scenarios. The AmpereOne A192-32X never consumed less than 101 Watts even in the brief periods of being idle... Meanwhile the single Intel / AMD server processors could idle in the 10~20'ish Watt range. The Ampere Altra Max M128-30 could sip as little as 21 Watts. But the AmpereOne didn't drop below 100 Watts.
The EPYC 9754 Bergamo had a similar power consumption overall to the AmpereOne A192-32X with a 225 Watt average and 397 Watt peak. The EPYC 9754 would consume just 10 Watts at idle.
The high CPU power use at idle also translated into much higher total AC server power consumption. The AmpereOne A192-32X Supermicro server consumed ~240 Watts at idle compared to ~106 Watts with Ampere Altra Max M128-30 on the Gigabyte 2U server. On average over the entire duration of benchmarks the AmpereOne A192-32X Supermicro ARS-211M-NR R13SPD server with A192-32X and 8 x 64GB DDR5-5200 memory was consuming 431 Watts with a recorded peak of 696 Watts. As stated earlier in the article, there weren't AC total power consumption measurements for EPYC and Xeon CPUs in this article due to using the reference servers there and all the vendors not recommending total power measurements from the reference platforms not to being as well tuned as those available from the OEMs/ODMs.
The AmpereOne performance for the most part was better than I had anticipated after this painfully slow ramp. We've been talking about AmpereOne for two years and now it's finally beginning to appear in the public cloud over at Oracle as GA, the first reviewer systems are out, and hopefully in the coming weeks/months will appear in-stock/available from more server providers. As of writing the Supermicro ARS-211M-NR R13SPD as used for this reviewer system isn't listed as readily available. Once the ramp-up for AmpereOne is there, it will be interesting to see the server pricing as well as how close the AmpereOne processors sell compared to their suggested pricing. At the $5555 USD suggested price for this 192-core 3.2GHz AmpereOne processor, it's much lower than the EPYC 9754 listed at around $8~9k from some Internet retailers. If that mid-$5k pricing holds for the AmpereOne A192-32X it can be quite compelling in a value sense if there ends up being robust availability of AmpereOne processors and the server/platform pricing isn't skewed higher. Meanwhile the Intel Xeon 6766E carries a list price of $10.2k but as of writing that Sierra Forest CPU can't be found in stock at any major Internet retailer.
Overall this was a better than expected first look at AmpereOne. It's great finally seeing AmpereOne hands-on and the performance being competitive. If AmpereOne had ramped up last year well before Intel Sierra Forest it would have looked extremely compelling with its 192 core part and launching to a very different landscape. Now it's appearing with Intel Sierra Forest having been "launched" since earlier this summer, Intel Granite Rapids coming up soon, and AMD EPYC Turin also due out in the coming months. It will be interesting to see if Ampere Computing can achieve robust availability on AmpereOne processors and platforms before Granite Rapids and EPYC Turin. Hopefully firmware updates can also help with AmpereOne reducing its idle CPU power usage.
So far Ampere Computing also hasn't announced any AmpereOne products below 96 cores but is said to continuing to sell Ampere Altra for years to come. This is a bit sad as beyond the cloud native focus, Ampere is the go-to solution for those wanting to build an ARM workstation. Linus Torvalds even loves his Ampere workstation as covered in various Phoronix articles. But Ampere Altra is years old, still relying on DDR4, etc. It would be great if there becomes an AmpereOne 32~64 core option for those wanting a modern AArch64 workstation. We'll see if anything materializes there following the AmpereOne server/cloud ramp.
Meanwhile to push things higher AmpereOne M with 12 channel DDR5 memory rather than 8 channels is expected to be out in Q4. Next year will be AmpereOne MX with up to 256 cores. Hopefully Ampere Computing manages to ramp these upcoming products faster to better compete with Intel Xeon 6 and AMD EPYC Turin.
Thanks to Ampere Computing for supplying this Supermicro ARS-211M-NR R13SPD review server with the AmpereOne A192-32X. Stay tuned for more AmpereOne A192-32X benchmarks and other follow-up articles while having my hands on this server the next few weeks.
If you enjoyed this article consider joining Phoronix Premium to view this site ad-free, multi-page articles on a single page, and other benefits. PayPal or Stripe tips are also graciously accepted. Thanks for your support.